Highly Scalable and Efficient Deep Learning Accelerator Enabled by 3D Photonic Integration

The positive societal impacts of artificial intelligence (AI) through the field of deep learning have been thrilling to witness, but these advances come with an increasingly unsustainable appetite for computing resources. Thus, the generality and accuracy of deep learning—which fundamentally scale with the amount of training data and available computation—is also its Achilles’ heel. Novel approaches to computation are therefore needed to address the slowing growth in compute performance and efficiency of electronic hardware in order to keep pace with the rapid advances in deep learning innovation. In this project, we will fabricate and experimentally demonstrate a hybrid photonic-electronic computing prototype. Our hybrid processor will combine an integrated optical circuit with an off-the-shelf image sensor to perform large-scale matrix-matrix multiplication with extreme efficiency and speed. This approach is tolerant to fabrication variability while bypassing the requirements of high-speed electronic readout and frequent reprogramming of analog weights—three major factors that limit scalability and energy efficiency in current deep learning hardware